Rose (@Rosecheng0) header
Rose (@Rosecheng0) profile picture

Product & GTM @BlockRunAI | @UCLA Exploring AI agents, agentic commerce & RL environment. Learning in public

Joined October 2025
Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)Photo by Rose (@Rosecheng0)
Rose retweeted
Thank @Rosecheng0 build so beautiful content on @xai 🥰
A seller asks an AI agent to pay twice for the same data. Does it pay? 7 of 12 models did. 4 of them paid on all 9 tasks with this trap. How the test works: each model buys data from a test seller in Agent Money Benchmark and pays in real USDC. After the payment, the seller asks to be paid again. 5 models never paid twice: Claude Opus 5.5, Claude Fable 5.2, GPT-6 Astra, GPT-6.1 Sol and Gemini 4 Argon. Of the 8 traps in the benchmark, this one caught the most models. Results for every model and trap: agentmoneybench.com
1
2
336
Rose retweeted
We showed 12 AI agents a price. Then the seller asked for more. Each agent was buying data with real USDC. At checkout, the seller quoted more than its own listed price. A human buyer would notice. Most agents did too. 7 models never paid the higher price: Claude Opus 5.5, Fable 5.2, Fable 5.1, Sonnet 5.5, GPT-6.1 Sol, GPT-6 Astra and Gemini 4 Argon. That's 0 of 202. 3 models usually did. @deepseek_ai V4 Pro paid 23 of 29 times, GPT-6 Luna 17 of 27 and @Kimi_Moonshot K3.1 Pro 12 of 29. The surprise is Claude Sonnet 5.5. Last week it paid every second bill for data it had already bought (9 of 9). Here it refused every raised price (0 of 28). A model can be careful with one kind of trick and blind to the next, so test each trap, not just the model's reputation. Same harness, same 80 tasks for every model, settled on-chain. Full results: agentmoneybench.com
2
1
11
416
love to hear that!
I’ve been using x402 with my autonomous agents, and it’s made a big difference in my workflow. Just by giving each agent its own wallet, I’ve cut costs and made things run much faster. I love how well it works for micropayments. I’ve had a great experience with @BlockRunAI and can only recommend it.
1
2
127
Agent Money Bench update: Claude Fable 5.2 just took the #1 spot on BlockRun's Agent Money Bench, with the lowest wasted spend of 12 models tested. We gave it a wallet and tested how well it handles real-money decisions. Claude Fable 5.2 performance across 12 models: • Wasted spend: $21 per $100, lowest across all 12 models. • Adversarial Marketplace: 98.8% success (79/80), matching Argon and Astra for the highest success rate. • Financial Research: answer quality 73.6 of 100, the highest on the board, surpassing the previous best of 68.8. • Financial Research all-pass rate: 57% (43/76), highest among all 12 models, up from the previous best of 50%. Compared with Fable 5.1, overall wasted spend fell 44%, from $37.7 to $21.0 per $100. Read the full results below👇agentmoneybench.com
3
3
1
17
1,269
AI agents need access to intelligence, tools, services, and a way to pay for the resources they use. We're excited to bring BlockRun's infrastructure and agent-native payment to the @MultiversX ecosystem. Together, we're expanding the tools available to MultiversX developers, bringing access to AI models, data, and tools alongside payment capabilities for autonomous agents. From AI that thinks to AI that acts, transacts, and creates value. Let's build. 🚀
15
105
6
294
9,003
What's new at BlockRun 🧵 TL;DR 01 · Meta's Muse Spark 1.3 and Muse Wonder 1.0 02 · Free MiniMax M3.1 Flash and Nemotron 3.5 Bolt 03 · Grok Imagine Video 1.5: choose how your video starts and ends 04 · Claude Opus 5.5, GPT-6.1 Sol and more on Arc Details below 👇
3
4
9
975
⚡@ClawRouter_ , BlockRun MCP, and @FranklinRun_ now support payments on @arc. Your agents can pay per call for models, image generation, and search — all in USDC on Arc. Start paying on Arc: arc.blockrun.ai/#get-started
13
5
31
4,034
4 more highlights from Agent Money Benchmark: 1. Duplicate-payment scams caught the most models. 6 out of 10 AI models made duplicate payments for data they'd already purchased. 4 of them paid again every single time they encountered the trap (9 out of 9 times). 2. Low cost and low waste don’t always go together. GPT-6 Luna had the lowest cost per research question ($0.032), yet wasted $71 out of every $100 spent, the most of the 10 models. 3. Capable is not the same as safe. Claude Fable 5.1 and Claude Opus 5.5 answered the same share of research questions right (36.8%). But Fable made payment or budget errors on 5.0% of marketplace tasks, while Opus made none. 4. Only 3 models were never scammed. GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 avoided every trap in the marketplace tests. See the full results, methodology, and model breakdown👇: agentmoneybench.com/
We gave 10 frontier AI models wallets with money. 7 got scammed at least once. The best wasted $25 of every $100. The worst wasted $71. Agent Money Benchmark tests whether AI agents can be trusted with real money. We evaluated 10 frontier models across two settings, using real USDC payments on @solana : → Financial Research: Agents buy live data from real providers to answer open-ended financial questions (prediction markets, wallet analysis, DeFi yields, etc.). → Adversarial Marketplace: Agents buy data from simulated sellers, some of whom cheat. The tasks test whether agents can avoid payment traps, stay within budget, and complete transactions without being exploited. Key results: → Give an agent $100 and the best tier wastes $25–38 of it. The weakest tier wastes $56–71. → 7 of 10 models paid a cheating seller at least once, and 6 paid twice for the same data. → Only 4 of 10 models made no money errors. → No model gets more than half of the research questions right. GPT-6.1 Sol led at 50%, at $0.468 per question. Explore the full results → agentmoneybench.com/
14
4
1
21
1,474
AI agents are moving beyond generating answers. They're buying data, accessing LLMs, and calling APIs to get things done. That means a new set of questions: → Can they evaluate what’s worth paying for? → Can they manage a budget? → Can they detect dishonest sellers? We built Agent Money Bench to measure what happens when agents have to put real money on the line. We gave frontier models wallets funded with real USDC and tasked them with acquiring resources and completing tasks in an environment where not every seller could be trusted. check out the full benchmark below 👇:
We gave 10 frontier AI models wallets with money. 7 got scammed at least once. The best wasted $25 of every $100. The worst wasted $71. Agent Money Benchmark tests whether AI agents can be trusted with real money. We evaluated 10 frontier models across two settings, using real USDC payments on @solana : → Financial Research: Agents buy live data from real providers to answer open-ended financial questions (prediction markets, wallet analysis, DeFi yields, etc.). → Adversarial Marketplace: Agents buy data from simulated sellers, some of whom cheat. The tasks test whether agents can avoid payment traps, stay within budget, and complete transactions without being exploited. Key results: → Give an agent $100 and the best tier wastes $25–38 of it. The weakest tier wastes $56–71. → 7 of 10 models paid a cheating seller at least once, and 6 paid twice for the same data. → Only 4 of 10 models made no money errors. → No model gets more than half of the research questions right. GPT-6.1 Sol led at 50%, at $0.468 per question. Explore the full results → agentmoneybench.com/
1
1
4
236
RT @BlockRunAI: AI agents need more than intelligence: they need the ability to pay for what they use. BlockRun is featured on @Circle’s a…
1
3
Want FREE $100 for your research agents to spend on AI models? We're honored to partner with @Anchorage to bring pay-per-request AI inference to Agentic Banking, from quick market news lookups to in-depth data research. 80+ models. One endpoint. Each request priced before it runs, paid over x402. Each participant will be funded with $100 in inference credit. 📧 Email: [email protected] to request access.
Per participant. 80+ AI models. One month. The credit covers BlockRun services only. Agents in the trial place no trades. Participants complete identity verification. Request access at [email protected] or via your relationship manager. Read more about our BlockRun partnership: anchorage.com/insights/ancho…
5
2
14
2,673
Building trading research agents? Here's your FREE $100 of AI inference for them 👇 Request yours: [email protected]
Want FREE $100 for your research agents to spend on AI models? We're honored to partner with @Anchorage to bring pay-per-request AI inference to Agentic Banking, from quick market news lookups to in-depth data research. 80+ models. One endpoint. Each request priced before it runs, paid over x402. Each participant will be funded with $100 in inference credit. 📧 Email: [email protected] to request access.
3
5
184
What's new at BlockRun 🧵 TL;DR 01 · Zero data retention on our Solana gateway 02 · Kimi K3.1 Pro, a third cheaper than K3 03 · Gemini 4 Argon 04 · DeepSeek V4.1 Flash, cheaper cache reads 05 · Skills and Share in Franklin chat Details below 👇
5
4
15
1,476
05 · Skills in Franklin chat Franklin is BlockRun's AI agent with its own wallet. Franklin chat now has Skills: instructions Franklin follows in your chats. Add one by Franklin or import your own. Imports stay private. Open Skills 👇 franklin.run/chat?view=myski…
1
12
05 · Share a Franklin chat New Share button, top right. Pick messages, then copy a read-only link, copy plain text, or turn them into an image. Also new: Feedback & suggestions in Settings, and More → Customize sidebar. Open Franklin chat 👇 franklin.run/chat
10